Re: [PATCH] Introduce atomic_long_t and asm-generic/atomic.h

[Andrew Morton <akpm@osdl.org>]

Christoph Lameter <clameter@engr.sgi.com> wrote:
>
> +#ifdef ATOMIC64_INIT
> +
> +#define ATOMIC_LONG_INIT(i)	ATOMIC64_INIT(i)
> +#define atomic_long_t		atomic64_t
> +#define atomic_long_read(v)	atomic64_read(v)
> +#define atomic_long_set(v,i)	atomic64_set(v,i)
> +#define atomic_long_inc(v)	atomic64_inc(v)
> +#define atomic_long_dec(v)	atomic64_dec(v)
> +#define atomic_long_add(i,v)	atomic64_add(i,v)
> +#define atomic_long_sub(i,v)	atomic64_sub(i,v)
> +
> +#else
> +
> +#define ATOMIC_LONG_INIT(i)	ATOMIC_INIT(i)
> +#define atomic_long_t		atomic_t
> +#define atomic_long_read(v)	atomic_read(v)
> +#define atomic_long_set(v,i)	atomic_set(v,i)
> +#define atomic_long_inc(v)	atomic_inc(v)
> +#define atomic_long_dec(v)	atomic_dec(v)
> +#define atomic_long_add(i,v)	atomic_add(i,v)
> +#define atomic_long_sub(i,v)	atomic_sub(i,v)
> +
> +#endif

It's sneaky, but it's not really good enough, IMO.

It assumes that sizeof(long) = sizeof(int) ifndef ATOMIC64_INIT.  Which is
true, but there are still problems.  For example, I'd reasonably expect this:

	printk("%ld", atomic_long_read(v));

to not produce a warning.  It may also lead to long*/int* warnings or build
errors.

Also, it kind-of assumes that each 64-bit arch uses `long' for its 64-bit
value.  sh64, for example, appears to use `long long'.


Perhaps all this can be fixed by filling the above macros with typecasts. 
Remember that typecasted lvals are illegal with gcc-4.x.

Or we bite the bullet and implement these guys in each arch...